Pith. sign in

REVIEW 1 major objections 5 minor 1 cited by

Rectified Flow models memorize training images most at the midpoint of their flow, and U-shaped timestep sampling reduces that risk without harming image quality.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-15 12:02 UTC pith:L2RNAOU4

load-bearing objection First practical MIA suite for Rectified Flow, a clear midpoint memorization peak with a usable U-shaped fix, and solid empirical gains; the exact LMMSE orthogonality argument is the softest formal link but not load-bearing. the 1 major comments →

arxiv 2603.13421 v2 pith:L2RNAOU4 submitted 2026-03-12 cs.LG cs.CV

Generalization and Memorization in Rectified Flow

Mingxing Rao , Daniel Moyer This is my paper
classification cs.LG cs.CV
keywords Rectified FlowFlow Matchingmembership inferencememorizationtimestep samplinginput complexitygeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Rectified Flow is a popular way to generate images by learning a straight path from noise to data. This paper asks how and when those models memorize their training set. The authors build membership-inference tests that detect whether a given image was in the training data, then correct those tests for the simple fact that easy-to-compress images look more familiar. With the corrected tests they show that, under ordinary uniform sampling of time, attack success is highest exactly halfway through the flow. They explain this by showing that at the midpoint a linear estimator becomes useless, so the network must use its nonlinear capacity on sample-specific details. Replacing uniform time sampling with a U-shaped (Symmetric Exponential) distribution keeps the model away from that vulnerable middle and measurably lowers memorization while leaving generative quality (FID) intact across three image datasets.

Core claim

Under standard uniform temporal training, a Rectified Flow model's susceptibility to membership inference strictly peaks at the integration midpoint t = 0.5; the expected memorization risk over continuous time is upper-bounded by the risk measured at that midpoint. Replacing uniform sampling by a Symmetric Exponential (U-shaped) distribution suppresses the peak and thereby reduces memorization while preserving generative fidelity.

What carries the argument

The complexity-calibrated Monte-Carlo statistic Tmc_cal = Tmc(x,t)/C(x), together with the LMMSE orthogonality argument that at t = 0.5 the state xt becomes uncorrelated with the target velocity, forcing the network off the linear baseline and maximizing sample-specific memorization.

Load-bearing premise

The claim that the peak sits exactly at the midpoint rests on second-moment orthogonality between the noisy state and the target velocity; if the data distribution or the learned velocity field break those moment assumptions, the peak need not stay at t = 0.5.

What would settle it

Train identical Rectified Flow models on the same data with uniform versus Symmetric-Exponential timestep sampling, measure Tmc_cal attack AUC and TPR@1%FPR at many discrete times, and check whether the uniform model still peaks at t = 0.5 while the U-shaped model shows a lower peak and comparable FID.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 5 minor

Summary. The paper studies memorization in Rectified Flow (RF) models via membership inference attacks (MIAs). It derives three test statistics from the Flow Matching and Conditional Flow Matching objectives (Tnaive, Tmc, and a complexity-calibrated Tmc_cal that normalizes by compressed bitrate C(x)), shows large gains from calibration (up to +15 AUC / +45 TPR@1%FPR), and reports that under uniform timestep sampling the MIA signal peaks at the integration midpoint t=0.5. This peak is linked to an LMMSE argument: at t=0.5 the state xt is orthogonal to the target velocity under second-moment assumptions, forcing the network off linear estimators. Replacing uniform sampling by a Symmetric Exponential (U-shaped) schedule is shown to suppress the peak while preserving FID on CIFAR-10, SVHN and TinyImageNet.

Significance. If the empirical pattern and mitigation hold, the work supplies the first non-trivial, RF-specific MIA and a practical temporal regularizer that improves the privacy–utility trade-off without DP-SGD-style noise. The progressive derivation of the three statistics from the training objectives, the explicit complexity calibration motivated by known likelihood biases, the consistent multi-dataset results (Table 2, Figs. 1–3, 8), the released code/checkpoints, and the closed-form LMMSE justification (Appendix E) are concrete strengths that make the claims falsifiable and reusable. The U-shaped schedule re-purposes an existing idea with a new memorization-theoretic rationale, which is a useful contribution for practitioners training large RF models.

major comments (1)
  1. Section 4 and Appendix E: the claim that susceptibility is strictly upper-bounded by the value at t=0.5 rests on the second-moment orthogonality argument that xt is statistically orthogonal to the target velocity exactly at the midpoint. Image data are high-dimensional and non-Gaussian; the paper should either (i) state the precise moment conditions under which the peak location is guaranteed, or (ii) replace the word “strictly” by “empirically / approximately” and add a short sensitivity check (e.g., non-zero mean or heavy-tailed synthetic data) so that the analytic upper-bound claim is not stronger than the derivation supports. The multi-dataset empirical peak and the U-shaped mitigation remain intact either way.
minor comments (5)
  1. Figure 1 / Table 2: report the number of Monte-Carlo samples N used for the main Tmc / Tmc_cal numbers (Fig. 5 shows the dependence but the primary tables do not).
  2. Section 3.2 and Appendix A: the complexity proxy C(x) is the compressed byte length; specify the exact compressor (PNG, JPEG quality, etc.) so that the calibration is reproducible.
  3. Notation: the same symbol T is used for both the test statistic and the time horizon; a subscript or different letter would avoid momentary confusion when reading Eqs. (5)–(10).
  4. Appendix B / Fig. 7: the claim that the ordinary training/validation loss gap is uninformative is important; a quantitative measure (e.g., AUC of the loss itself) would make the contrast with Tmc_cal sharper.
  5. Typos / formatting: “Schocastic” (p. 4), “Recitified” (Appendix B title), and several mid-sentence page breaks / missing figure captions in the supplied PDF should be cleaned.

Circularity Check

0 steps flagged

No significant circularity: MIA statistics derive directly from FM/CFM objectives, midpoint peak is an empirical observation independently justified by LMMSE orthogonality under stated second-moment assumptions, and U-shaped sampling is an external schedule re-purposed without self-definition.

full rationale

The three test statistics are obtained by rewriting the Flow Matching and Conditional Flow Matching losses (Eqs. 3-4) into evaluable residuals (Tnaive, Tmc) and then dividing by an independent compression-bitrate complexity measure C(x); none of these quantities is defined in terms of the membership labels or the peak location they later measure. The claim that susceptibility peaks at t=0.5 is first reported as an empirical pattern under uniform sampling (Fig. 1) and only afterwards given a separate analytic justification via the LMMSE estimator (Section 4 and Appendix E), which shows statistical orthogonality of xt and the target velocity at the midpoint under global mean/covariance assumptions; the derivation does not presuppose the peak. The Symmetric Exponential schedule is taken from Lee et al. (2024) and merely re-used; no parameter is fitted to the MIA curves and then re-predicted. Self-citations to the authors' prior diffusion-MIA papers appear only as background and are not load-bearing for any equation or bound in the present work. Consequently the derivation chain is self-contained against its own inputs.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claims rest on the standard Flow Matching / Conditional Flow Matching objectives, the known equivalence of their gradients, the use of compressed bitrate as a proxy for Kolmogorov complexity, second-moment assumptions that make xt orthogonal to the target velocity at t=0.5, and the empirical choice of Monte-Carlo sample size and concentration parameter alpha. No new physical entities are postulated.

free parameters (3)
  • number of Monte Carlo samples N (#mc) = 5 (main results)
    Used to approximate the expectation in T_mc and T_mc_cal; set to 5 in the main tables; performance improves with larger N but is treated as a free computational choice.
  • Symmetric Exponential concentration alpha
    Controls how strongly the U-shaped schedule avoids the midpoint; varied experimentally (blue/green/orange curves) and chosen to trade off memorization versus FID.
  • compressed bitrate C(x) as complexity proxy
    Used to normalize T_mc; the precise compressor and byte-length definition are implementation choices that affect the numerical scale of the calibrated statistic.
axioms (4)
  • domain assumption Gradient of the marginal Flow Matching objective equals the gradient of the Conditional Flow Matching objective, so optimizing CFM yields the same velocity field as FM.
    Invoked in Section 3 to justify deriving the Monte-Carlo test statistic from the practical CFM loss (Lipman et al.).
  • ad hoc to paper At t=0.5 the noisy state xt is statistically orthogonal to the target velocity under second-moment assumptions on the data and noise, forcing the network away from the LMMSE linear estimator.
    Core of the mathematical justification for the midpoint peak (Section 4 and Appendix E).
  • domain assumption Compressed file size (bitrate) is a sufficient proxy for the intrinsic spatial complexity that biases likelihood and reconstruction scores.
    Taken from Serrà et al. and Nalisnick et al. and used to define the calibrated statistic T_mc_cal.
  • domain assumption Membership-inference success is a valid quantitative proxy for sample-specific memorization in continuous-time generative models.
    Standard in the MIA literature (Carlini et al.) and adopted throughout the paper.

pith-pipeline@v1.1.0-grok45 · 16251 in / 2909 out tokens · 28297 ms · 2026-07-15T12:02:35.343344+00:00 · methodology

0 comments
read the original abstract

Generative models based on the Flow Matching objective, particularly Rectified Flow, have emerged as a dominant paradigm for efficient, high-fidelity image synthesis. However, while existing research heavily prioritizes generation quality and architectural scaling, the underlying dynamics of how RF models memorize training data remain largely underexplored. In this paper, we systematically investigate the memorization behaviors of RF through the test statistics of Membership Inference Attacks (MIA). We progressively formulate three test statistics, culminating in a complexity-calibrated metric that successfully decouples intrinsic image spatial complexity from genuine memorization signals. This calibration yields a significant performance surge -- boosting attack AUC by up to 15% and the privacy-critical TPR@1%FPR metric by up to 45% -- establishing the first non-trivial MIA specifically tailored for RF. Leveraging these refined metrics, we uncover a distinct temporal pattern: under standard uniform temporal training, a model's susceptibility to MIA strictly peaks at the integration midpoint, a phenomenon we justify via the network's forced deviation from linear approximations. Finally, we demonstrate that substituting uniform timestep sampling with a Symmetric Exponential (U-shaped) distribution effectively minimizes exposure to vulnerable intermediate timesteps. Extensive evaluations across three datasets confirm that this temporal regularization suppresses memorization while preserving generative fidelity.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Theoretical Analysis of Memory and Overfitting Phenomena in Stochastic Interpolation Models

    cs.LG 2026-06 unverdicted novelty 5.0

    In the oracle continuous-time setting, stochastic interpolation models recover training samples exactly, with deviations controlled by discretization and estimation errors, leading to theoretical definitions of overfi...

Reference graph

Works this paper leans on

50 extracted references · 20 linked inside Pith · cited by 1 Pith paper

  1. [1]

    In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security

    Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308–318 (2016)

    2016

  2. [2]

    arXiv preprint arXiv:2209.15571 (2022)

    Albergo, M.S., Vanden-Eijnden, E.: Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571 (2022)

    Pith/arXiv arXiv 2022

  3. [3]

    arXiv preprint arXiv:2505.17638 (2025)

    Bonnaire, T., Urfin, R., Biroli, G., Mézard, M.: Why diffusion models don’t memorize: The role of implicit dynamical regularization in training. arXiv preprint arXiv:2505.17638 (2025)

    arXiv 2025

  4. [4]

    Advances in neural information processing systems34, 27916–27928 (2021)

    van den Burg, G., Williams, C.: On memorization in probabilistic deep generative models. Advances in neural information processing systems34, 27916–27928 (2021)

    2021

  5. [5]

    In: 2022 IEEE symposium on security and privacy (SP)

    Carlini, N., Chien, S., Nasr, M., Song, S., Terzis, A., Tramer, F.: Member- ship inference attacks from first principles. In: 2022 IEEE symposium on security and privacy (SP). pp. 1897–1914. IEEE (2022)

    2022

  6. [6]

    In: 32nd USENIX security symposium (USENIX Security 23)

    Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX security symposium (USENIX Security 23). pp. 5253–5270 (2023)

    2023

  7. [7]

    In: 2009 IEEE conference on com- puter vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on com- puter vision and pattern recognition. pp. 248–255. Ieee (2009)

    2009

  8. [8]

    arXiv preprint arXiv:2210.09929 (2022)

    Dockhorn, T., Cao, T., Vahdat, A., Kreis, K.: Differentially private diffusion models. arXiv preprint arXiv:2210.09929 (2022)

    Pith/arXiv arXiv 2022

  9. [9]

    Duan, J., Kong, F., Wang, S., Shi, X., Xu, K.: Are diffusion models vul- nerable to membership inference attacks? In: International Conference on Machine Learning. pp. 8717–8730. PMLR (2023)

    2023

  10. [10]

    In: Forty-first international confer- ence on machine learning (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transform- ers for high-resolution image synthesis. In: Forty-first international confer- ence on machine learning (2024)

    2024

  11. [11]

    In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security

    Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322–1333 (2015)

    2015

  12. [12]

    arXiv preprint arXiv:2308.12143 (2023)

    Fu,W.,Wang,H.,Gao,C.,Liu,G.,Li,Y.,Jiang,T.:Aprobabilisticfluctua- tion based membership inference attack for diffusion models. arXiv preprint arXiv:2308.12143 (2023)

    Pith/arXiv arXiv 2023

  13. [13]

    arXiv preprint arXiv:2310.17432 (2023) Generalization and Memorization in Rectified Flow 17

    Goodier, J., Campbell, N.D.: Likelihood-based out-of-distribution de- tection with denoising diffusion probabilistic models. arXiv preprint arXiv:2310.17432 (2023) Generalization and Memorization in Rectified Flow 17

    Pith/arXiv arXiv 2023

  14. [14]

    arXiv preprint arXiv:1810.01367 (2018)

    Grathwohl, W., Chen, R.T., Bettencourt, J., Sutskever, I., Duvenaud, D.: Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367 (2018)

    Pith/arXiv arXiv 2018

  15. [15]

    arXiv preprint arXiv:2310.02664 (2023)

    Gu, X., Du, C., Pang, T., Li, C., Lin, M., Wang, Y.: On memorization in diffusion models. arXiv preprint arXiv:2310.02664 (2023)

    Pith/arXiv arXiv 2023

  16. [16]

    arXiv preprint arXiv:1705.07663 (2017)

    Hayes, J., Melis, L., Danezis, G., De Cristofaro, E.: Logan: Membership in- ference attacks against generative models. arXiv preprint arXiv:1705.07663 (2017)

    Pith/arXiv arXiv 2017

  17. [17]

    Ad- vances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Ad- vances in neural information processing systems33, 6840–6851 (2020)

    2020

  18. [18]

    arXiv preprint arXiv:2310.02557 (2023)

    Kadkhodaie, Z., Guth, F., Simoncelli, E.P., Mallat, S.: Generalization in diffusion models arises from geometry-adaptive harmonic representations. arXiv preprint arXiv:2310.02557 (2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Advances in neural information process- ing systems35, 26565–26577 (2022)

    Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in neural information process- ing systems35, 26565–26577 (2022)

    2022

  20. [20]

    arXiv preprint arXiv:1312.6114 (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    Pith/arXiv arXiv 2013

  21. [21]

    Advances in neural information processing systems31(2018)

    Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 con- volutions. Advances in neural information processing systems31(2018)

    2018

  22. [22]

    arXiv preprint arXiv:2305.18355 (2023)

    Kong, F., Duan, J., Ma, R., Shen, H., Zhu, X., Shi, X., Xu, K.: An efficient membership inference attack for the diffusion model by proximal initializa- tion. arXiv preprint arXiv:2305.18355 (2023)

    Pith/arXiv arXiv 2023

  23. [23]

    Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)

    2009

  24. [24]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024)

    2024

  25. [25]

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Müller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: Flux.1 kontext: Flow matching for in-context image gener- ation and editing in latent space (2025),https://a...

    2025

  26. [26]

    Lee,S., Lin,Z.,Fanti,G.:Improvingthetrainingofrectifiedflows.Advances in neural information processing systems37, 63082–63109 (2024)

    2024

  27. [27]

    Li, M., Vitányi, P., et al.: An introduction to Kolmogorov complexity and its applications, vol. 3. Springer (2008)

    2008

  28. [28]

    Advances in Neural Information Processing Systems36, 2097– 2127 (2023)

    Li, P., Li, Z., Zhang, H., Bian, J.: On the generalization properties of diffu- sion models. Advances in Neural Information Processing Systems36, 2097– 2127 (2023)

    2097

  29. [29]

    Advances in neural information processing systems37, 57499–57538 (2024)

    Li, X., Dai, Y., Qu, Q.: Understanding generalizability of diffusion mod- els requires rethinking the hidden gaussian structure. Advances in neural information processing systems37, 57499–57538 (2024)

    2024

  30. [30]

    arXiv preprint arXiv:2210.02747 (2022)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

    Pith/arXiv arXiv 2022

  31. [31]

    arXiv preprint arXiv:2412.06264 (2024) 18 M

    Lipman, Y., Havasi, M., Holderrieth, P., Shaul, N., Le, M., Karrer, B., Chen, R.T., Lopez-Paz, D., Ben-Hamu, H., Gat, I.: Flow matching guide and code. arXiv preprint arXiv:2412.06264 (2024) 18 M. Rao and D. Moyer

    Pith/arXiv arXiv 2024

  32. [32]

    arXiv preprint arXiv:2209.03003 (2022)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022)

    Pith/arXiv arXiv 2022

  33. [33]

    arXiv preprint arXiv:2509.09672 (2025)

    Lukoianov, A., Yuan, C., Solomon, J., Sitzmann, V.: Locality in image dif- fusion models emerges from data statistics. arXiv preprint arXiv:2509.09672 (2025)

    arXiv 2025

  34. [34]

    Nalisnick, E., Matsukawa, A., Teh, Y.W., Gorur, D., Lakshminarayanan, B.: Do deep generative models know what they don’t know? arXiv preprint arXiv:1810.09136 (2018)

    Pith/arXiv arXiv 2018

  35. [35]

    In: NIPS workshop on deep learning and unsupervised feature learning

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y., et al.: Reading digits in natural images with unsupervised feature learning. In: NIPS workshop on deep learning and unsupervised feature learning. vol. 2011, p. 4. Granada (2011)

    2011

  36. [36]

    arXiv preprint arXiv:2304.14772 (2023)

    Pooladian, A.A., Ben-Hamu, H., Domingo-Enrich, C., Amos, B., Lipman, Y., Chen, R.T.: Multisample flow matching: Straightening flows with mini- batch couplings. arXiv preprint arXiv:2304.14772 (2023)

    Pith/arXiv arXiv 2023

  37. [37]

    arXiv preprint arXiv:2511.20592 (2025)

    Rao, M., Qu, B., Moyer, D.: Latent diffusion inversion requires understand- ing the latent space. arXiv preprint arXiv:2511.20592 (2025)

    arXiv 2025

  38. [38]

    arXiv preprint arXiv:2509.25003 (2025)

    Rao, M., Qu, B., Moyer, D.: Score-based membership inference on diffusion models. arXiv preprint arXiv:2509.25003 (2025)

    Pith/arXiv arXiv 2025

  39. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High- resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  40. [40]

    arXiv preprint arXiv:1909.11480 (2019)

    Serrà, J., Álvarez, D., Gómez, V., Slizovskaia, O., Núñez, J.F., Luque, J.: Input complexity and out-of-distribution detection with likelihood-based generative models. arXiv preprint arXiv:1909.11480 (2019)

    Pith/arXiv arXiv 1909

  41. [41]

    Advances in neural information processing systems32 (2019)

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems32 (2019)

    2019

  42. [42]

    arXiv preprint arXiv:2011.13456 (2020)

    Song,Y.,Sohl-Dickstein,J.,Kingma,D.P.,Kumar,A.,Ermon,S.,Poole,B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

    Pith/arXiv arXiv 2011

  43. [43]

    arXiv preprint arXiv:2302.00482 (2023)

    Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., Bengio, Y.: Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482 (2023)

    Pith/arXiv arXiv 2023

  44. [44]

    arXiv preprint arXiv:2011.11660 (2020)

    Tramer, F., Boneh, D.: Differentially private learning needs better features (or much more data). arXiv preprint arXiv:2011.11660 (2020)

    Pith/arXiv arXiv 2011

  45. [45]

    Van Den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neuralnetworks.In:Internationalconferenceonmachinelearning.pp.1747–

  46. [46]

    Research in the Mathematical Sciences9(1), 8 (2022)

    Yang, H., E, W.: Generalization error of gan from the discriminator’s per- spective. Research in the Mathematical Sciences9(1), 8 (2022)

    2022

  47. [47]

    In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition

    Yang, X., Zhou, D., Feng, J., Wang, X.: Diffusion probabilistic model made slim. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 22552–22562 (2023) Generalization and Memorization in Rectified Flow 19

    2023

  48. [48]

    In: ICML 2023 workshop on struc- tured probabilistic inference{\&}generative modeling (2023)

    Yoon, T., Choi, J.Y., Kwon, S., Ryu, E.K.: Diffusion probabilistic models generalize when they fail to memorize. In: ICML 2023 workshop on struc- tured probabilistic inference{\&}generative modeling (2023)

    2023

  49. [49]

    Advances in Neural Information Processing Systems37, 74122–74146 (2024) 20 M

    Zhai, S., Chen, H., Dong, Y., Li, J., Shen, Q., Gao, Y., Su, H., Liu, Y.: Membershipinferenceontext-to-imagediffusionmodelsviaconditionallike- lihood discrepancy. Advances in Neural Information Processing Systems37, 74122–74146 (2024) 20 M. Rao and D. Moyer 7 Appendix A Input Complexity Bias in Likelihood Estimation For a trainedvθ on RF. The evolution of...

    2024

  50. [50]

    B Training/val Loss Gap in Recitified Flow As shown in Fig

    The corresponding compressed rate (bytes) is shown for each image. B Training/val Loss Gap in Recitified Flow As shown in Fig. 7, the original RF loss exhibits negligible separability be- tweenD train andD val, suggested by the near-complete overlap (indicated by the purple region) of their±1standard deviation intervals. This renders it largely 22 M. Rao ...

    2000